Thermal power generating unit low-frequency oscillation fault diagnosis method and system based on multi-source data analysis

Abstract

The invention discloses a thermal power generating unit low-frequency oscillation fault diagnosis method and system based on multi-source data analysis, and mainly relates to the technical field of thermal power generating unit fault diagnosis. Comprising the steps that a multi-source signal acquisition network is constructed, multi-source unit operation sensing of a preset monitoring window is executed on a target thermal power unit, and a speed regulator oil pressure pulsation sequence, a shaft system torsional vibration angular displacement sequence, a boiler heat storage coefficient sequence and a power grid power angle sequence are obtained; cascade time-frequency analysis is carried out; performing joint frequency matching and mode identification to obtain a plurality of abnormal low-frequency oscillation modes; and performing fault scene diagnosis on the plurality of abnormal low-frequency oscillation modes to obtain a target diagnosis result. The low-frequency oscillation fault diagnosis method has the beneficial effects that the technical problem of high error rate of low-frequency oscillation fault diagnosis caused by lack of reliable analysis of signals in the prior art is solved, and the technical effect of improving the fault diagnosis accuracy is achieved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for thermal power units, specifically to a method and system for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis. Background Technology

[0002] When thermal power units operate under conditions of deep peak shaving and large-scale grid integration of new energy sources, low-frequency oscillations (typically located in the 0.1–2 Hz frequency band) are prone to occur between the speed control system, shaft system, and power grid. Traditional oscillation monitoring systems often rely on single sensor signals, such as unit speed or power fluctuations. This approach cannot distinguish oscillations caused by various mechanical, thermal, and electrical factors. For example, mechanical torsional vibration, thermal inertia fluctuations, and power grid power angle oscillations are often mixed together in the oscillation signal, and a single signal source cannot accurately trace and distinguish the root cause of these oscillations.

[0003] Current spectrum analysis, while providing signal strength distributions across different frequency bands, fails to effectively capture the time-varying characteristics of oscillation modes when processing oscillating signals. Although spectrum analysis can identify the presence of oscillations, its timeliness is poor, typically exhibiting a warning delay of over 10 seconds, far exceeding the possible intervention window. Therefore, traditional monitoring systems suffer from signal lag and response delays, particularly in complex power systems and generating units, making it difficult to capture instantaneous signal changes during oscillations. This delayed response severely impacts system intervention and control effectiveness, easily leading to the spread of oscillation problems and consequently affecting grid stability and the safe operation of generating units. This results in low fault diagnosis accuracy, difficulty in precisely locating problems, and an inability to dynamically identify complex oscillation mechanisms. Summary of the Invention

[0004] This application provides a method and system for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis, which is used to address the technical problem of high error rates in low-frequency oscillation fault diagnosis due to the lack of reliable signal analysis in the prior art.

[0005] In view of the above problems, this application provides a method and system for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis.

[0006] The first aspect of this application provides a method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis, the method comprising: A multi-source signal acquisition network is constructed to perform multi-source unit operation sensing on the target thermal power unit within a preset monitoring window, obtaining governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and grid power angle sequence. Cascaded time-frequency analysis is performed on these sequences to obtain multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple grid power angle oscillation parameters. Joint frequency matching and mode identification are performed on these parameters to obtain multiple abnormal low-frequency oscillation modes. A large model is used to perform fault scenario-based diagnosis on each of these abnormal low-frequency oscillation modes to obtain target diagnosis results.

[0007] Furthermore, a multi-source signal acquisition network is constructed, including: deploying high-frequency pressure sensors at the inlet and outlet of the target thermal power unit; circumferentially arranging four sets of laser Doppler vibration meters at the high- and medium-pressure rotor of the target thermal power unit; installing steam drum pressure sensors and steam flow meters on the steam drum of the target thermal power unit; connecting a PMU device to synchronously capture the grid connection point frequency and power angle of the target thermal power unit; and constructing the multi-source signal acquisition network based on the high-frequency pressure sensor, the four sets of laser Doppler vibration meters, the steam drum pressure sensor, the steam flow meter, and the PMU device.

[0008] Furthermore, cascaded time-frequency analysis is performed on the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence to obtain multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters. This includes: performing multi-scale wavelet decomposition on the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence to filter out high-frequency random disturbances, and performing modal decomposition to obtain multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters. The intrinsic modal components of governor oil pressure pulsation, multiple shaft torsional angular displacement, multiple boiler heat storage coefficient, and multiple power grid power angle were analyzed using Prony analysis. This yielded multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters.

[0009] Furthermore, joint frequency matching and mode identification are performed on the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters to obtain multiple abnormal low-frequency oscillation modes. This includes: obtaining multiple frequency oscillation interest intervals; mapping the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters to obtain multiple candidate joint oscillation modes; performing scenario-based consistency verification on the multiple candidate joint oscillation modes; and adding them to the multiple abnormal low-frequency oscillation modes if the verification passes.

[0010] Furthermore, a scenario-based consistency check is performed on the multiple candidate joint oscillation modes. If the check passes, the modes are added to the multiple abnormal low-frequency modes. This includes: obtaining a set of historical abnormal low-frequency modes; aggregating the historical abnormal low-frequency modes by similar types to obtain multiple aggregated historical abnormal low-frequency mode sets; performing centralized representative screening on the multiple aggregated historical abnormal low-frequency mode sets to determine multiple screened historical abnormal low-frequency modes; and matching the multiple candidate joint oscillation modes with the multiple screened historical abnormal low-frequency modes. If there is a matching result with a matching degree greater than or equal to a preset matching degree threshold, the check passes; if there is no matching result with a matching degree greater than or equal to the preset matching degree threshold, the check fails.

[0011] Furthermore, the multiple aggregated historical low-frequency abnormal modes are subjected to centralized representative screening to determine multiple screened historical low-frequency abnormal modes. This includes: averaging the multiple aggregated historical low-frequency abnormal modes to determine the average of the multiple aggregated historical low-frequency abnormal modes; using the average of the multiple aggregated historical low-frequency abnormal modes as multiple initial centers, searching the multiple aggregated historical low-frequency abnormal modes according to a preset update bandwidth to determine multiple primary update centers; comparing the multiple primary update centers with the multiple initial centers to determine multiple first-stage centers; again searching the multiple aggregated historical low-frequency abnormal modes according to the preset update bandwidth to determine multiple secondary update centers, comparing them with the multiple first-stage centers to determine secondary update centers, and continuing the screening based on the comparison results until a preset screening stop condition is met to obtain multiple screened historical low-frequency abnormal modes.

[0012] Furthermore, a centralized representative comparison is performed on the plurality of primary update centers and the plurality of initial centers to determine a plurality of first-stage centers, including: calculating the nearest neighbor tendency coefficients of the plurality of primary update centers and the plurality of initial centers respectively to obtain the nearest neighbor tendency coefficients of the plurality of primary update centers and the plurality of initial centers; performing mapping comparisons on the nearest neighbor tendency coefficients of the plurality of primary update centers and the plurality of initial centers respectively, and taking the center corresponding to the larger value in the mapping comparison results as a plurality of first-stage centers.

[0013] Furthermore, the preset screening stopping condition is that the number of updates meets a preset maximum number of updates and / or the difference between the two nearest neighbor tendency coefficients of the mapping comparison is less than or equal to a preset difference. Further, a large model is used to perform fault scenario-based diagnosis on the multiple abnormal low-frequency oscillation modes to obtain the target diagnosis result, including: acquiring multiple sample abnormal low-frequency oscillation modes and multiple sample fault diagnosis results as training data; performing supervised training on a framework built based on a feedforward neural network to obtain a large model; using the large model to identify the multiple abnormal low-frequency oscillation modes to obtain multiple fault diagnosis results; and summarizing the multiple fault diagnosis results to obtain the target diagnosis result.

[0014] A second aspect of this application provides a low-frequency oscillation fault diagnosis system for thermal power units based on multi-source data analysis, the system comprising: The system comprises the following modules: a perception module for constructing a multi-source signal acquisition network; a cascaded time-frequency analysis module for performing multi-source unit operation perception on the target thermal power unit within a preset monitoring window; a cascaded time-frequency analysis module for performing cascaded time-frequency analysis on the governor oil pressure pulsation sequence, shaft torsional angular displacement sequence, boiler heat storage coefficient sequence, and grid power angle sequence; a modal identification module for performing joint frequency matching and modal identification on the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple grid power angle oscillation parameters; and a diagnostic module for performing fault scenario-based diagnosis on the multiple abnormal low-frequency oscillation modes using a large model to obtain target diagnostic results.

[0015] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application constructs a multi-source signal acquisition network to perform multi-source unit operation sensing on the target thermal power unit within a preset monitoring window, obtaining governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and grid power angle sequence. It then performs cascaded time-frequency analysis on these sequences to obtain multiple governor oil pressure pulsation oscillation parameters, shaft torsional vibration angular displacement oscillation parameters, boiler heat storage coefficient oscillation parameters, and grid power angle oscillation parameters. Finally, it performs joint frequency matching and mode identification on these parameters to obtain multiple abnormal low-frequency oscillation modes. Finally, it performs fault scenario-based diagnosis on each of these abnormal low-frequency oscillation modes to obtain the target diagnosis results. This achieves the technical effect of improving the accuracy of fault diagnosis. Attached Figure Description

[0016] Appendix Figure 1 This is a schematic diagram of the process for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis, provided in an embodiment of the present invention.

[0017] Appendix Figure 2 This is a schematic diagram of the structure of a low-frequency oscillation fault diagnosis system for thermal power units based on multi-source data analysis, provided in an embodiment of the present invention.

[0018] The labels shown in the attached diagram: The system includes a perception module 11, a cascaded time-frequency analysis module 12, a modal recognition module 13, and a diagnostic module 14. Detailed Implementation

[0019] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims. It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products, or devices.

[0020] Example 1, as shown in the appendix Figure 1 As shown, this application provides a method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis, wherein the method includes: Step S100: Construct a multi-source signal acquisition network to perform multi-source unit operation sensing on the target thermal power unit through a preset monitoring window, and obtain the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence and power grid power angle sequence; Furthermore, in constructing a multi-source signal acquisition network, step S100 of this embodiment also includes: High-frequency pressure sensors are deployed at the inlet and outlet of the target thermal power unit; Four sets of laser Doppler vibration meters were arranged circumferentially at the high- and intermediate-pressure rotor of the target thermal power unit. Install steam drum pressure sensors and steam flow meters on the steam drum of the target thermal power unit; The PMU device is connected to synchronously capture the grid connection point frequency and power angle of the target thermal power unit; The multi-source signal acquisition network is constructed based on a high-frequency pressure sensor, four sets of laser Doppler vibration meters, a steam drum pressure sensor, a steam flow meter, and a PMU device.

[0021] It should be noted that a multi-source signal acquisition network refers to a monitoring system integrating multiple sensors and synchronous acquisition devices, capable of simultaneously sensing the mechanical, thermal, and electrical operating status of a thermal power unit. The governor oil pressure pulsation sequence reflects the dynamic pressure changes in the hydraulic actuator of the speed control system, used to analyze servo control characteristics and hydraulic oscillations. The shaft torsional vibration angular displacement sequence represents the angular displacement change of the rotor in the torsional direction, reflecting the mechanical inertial response. The boiler heat storage coefficient sequence is a thermodynamic parameter calculated from the steam drum pressure and steam flow rate, used to characterize the boiler's thermal energy storage and thermal inertia changes. The power grid power angle sequence is acquired by a synchronous phasor measurement unit (PMU), representing the power phase relationship between the unit and the grid, and is a key indicator reflecting electrical stability. The preset monitoring window is a time range for data acquisition pre-set by those skilled in the art, such as 10 seconds, 30 seconds, etc.

[0022] Step S200: Traverse the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence and power grid power angle sequence to perform cascaded time-frequency analysis, and obtain multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters; Furthermore, by cascading time-frequency analysis of the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence, multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters are obtained. In this embodiment, step S200 further includes: The governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence are traversed and multi-scale wavelet decomposition is performed to filter out high-frequency random disturbances. Modal decomposition is then performed to obtain multiple governor oil pressure pulsation eigenmode components, multiple shaft torsional vibration angular displacement eigenmode components, multiple boiler heat storage coefficient eigenmode components, and multiple power grid power angle eigenmode components. Prony analysis was performed on multiple governor hydraulic pressure pulsation intrinsic mode components, multiple shaft torsional vibration angular displacement intrinsic mode components, multiple boiler heat storage coefficient intrinsic mode components, and multiple power grid power angle intrinsic mode components to obtain multiple governor hydraulic pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters.

[0023] In one possible embodiment, multi-scale wavelet decomposition is first performed on the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence, respectively. By selecting an appropriate wavelet basis, such as db4 or sym8, the complex mixed signal is decomposed into sub-signals of different frequency bands. The high-frequency layer, such as >5Hz, mainly represents random disturbances and servo responses; the mid-to-low frequency layer, such as 0.1–2Hz, reflects the inherent oscillations of the shaft system and hydraulic system; and the extremely low-frequency layer, such as <0.1Hz, reflects the slow variation process of boiler heat storage and power grid power angle. This step effectively removes noise interference and retains effective oscillation energy in the range of 0.05–10Hz.

[0024] Subsequently, modal decomposition, such as EMD / VMD, is performed on the effective components within each frequency band to obtain multiple intrinsic modal components. Each mode corresponds to a stable oscillation behavior in the signal, such as: servo oscillation of governor oil pressure, torsional inertial oscillation of shaft system, thermal inertial feedback fluctuation of boiler heat storage, and synchronous power swing of power grid angle. Prony analysis is applied to each modal component to establish a discrete exponential model. The damping factor and angular frequency are obtained through least squares fitting, and the frequency, damping ratio, and energy decay rate of each oscillation are calculated.

[0025] By analyzing the signals of oil pressure, torsional vibration, heat storage, and power angle layer by layer in the time-frequency-modal dimensions, the oscillation characteristics of different physical subsystems can be identified synchronously, providing a foundation for subsequent joint frequency matching and modal identification.

[0026] Step S300: Perform joint frequency matching and mode identification on the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters to obtain multiple abnormal low-frequency oscillation modes; Furthermore, joint frequency matching and mode identification are performed on the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters to obtain multiple abnormal low-frequency oscillation modes. In this embodiment, step S300 further includes: Multiple frequency oscillation interest intervals are obtained, and the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters are mapped to obtain multiple candidate joint oscillation modes. A scenario-based consistency check is performed on the multiple candidate joint oscillation modes. If the check passes, the multiple abnormal low-frequency oscillation modes are added.

[0027] In one possible embodiment, joint frequency matching refers to finding modal components with similar frequency, phase, or damping characteristics among the oscillation parameters extracted from signal sources with different oil pressures, torsional vibrations, heat storage, and power angles, in order to identify transdomain resonance phenomena that may belong to the same physical oscillation mechanism.

[0028] Multiple frequency oscillation interest intervals are key frequency bands preset by those skilled in the art based on the historical operating characteristics of the unit and statistical experience of power system oscillations. These typically fall within different frequency ranges of 0.1–2 Hz, and the main low-frequency dynamic characteristics of the corresponding thermal power unit are set within these intervals. Candidate joint oscillation modes refer to combinations of similar frequencies and stable phase relationships appearing in multiple signals, representing potential multi-physics resonance modes. Scenario-based consistency verification is used to confirm whether the mode is consistent with known mechanisms, thereby filtering out accidental frequency overlaps or noise pseudo-modes.

[0029] Preferably, based on historical modal samples from a historical database of abnormal low-frequency oscillations, the frequency distribution, energy coupling direction, and time delay characteristics of the current mode are compared with those of historical modes to determine their consistency. For example, if oil pressure fluctuations lead shaft torsional vibrations and both increase in energy synchronously at the 0.3Hz frequency point, it can be identified as a hydraulic-mechanical coupled mode. Alternatively, if changes in boiler heat storage lag behind power angle oscillations, it may be an electrical disturbance feedback-type thermal inertial oscillation. Through this semantic and data-driven comparison mechanism, invalid modes are filtered out, retaining only physically consistent abnormal low-frequency oscillation modes.

[0030] By establishing a common-mode identification mechanism for cross-domain oscillations, that is, by achieving coupled identification of signals from the mechanical, thermal, and electrical domains through joint matching, the limitations of traditional single-signal spectrum analysis can be overcome, and accurate modeling and identification of complex low-frequency resonance phenomena can be achieved.

[0031] Furthermore, a scenario-based consistency check is performed on the multiple candidate joint oscillation modes. If the check passes, the modes are added to the multiple abnormal low-frequency modes. Step S300 in this embodiment further includes: Obtain a set of historical abnormal low-frequency modes, and perform similar aggregation on the set of historical abnormal low-frequency modes to obtain multiple aggregated sets of historical abnormal low-frequency modes; The multiple aggregated historical abnormal low-frequency modes are centrally and representatively screened to determine multiple screened historical abnormal low-frequency modes. The candidate joint oscillation modes are matched with the screening historical abnormal low-frequency modes respectively. When there is a matching result with a matching degree greater than or equal to the preset matching degree threshold, the verification is passed. If there is no matching result with a matching degree greater than or equal to the preset matching degree threshold, the verification fails.

[0032] It should be noted that the historical abnormal low-frequency mode set refers to a typical low-frequency oscillation feature sample library accumulated through long-term unit operation monitoring or large model learning. Feature clustering algorithms, such as the K-means algorithm, are used to aggregate these mode sets, assigning modes with similar frequencies, damping ratios, and energy coupling directions to the same cluster, thus obtaining multiple aggregated historical abnormal low-frequency mode sets. Furthermore, a centralized representativeness screening is performed on the aggregated mode sets. By calculating the central feature of each cluster and performing multiple rounds of retrieval and updating based on the update bandwidth, i.e., the allowable modal feature offset range, the most representative modal samples are selected, resulting in multiple screened historical abnormal low-frequency modes.

[0033] Furthermore, using cosine similarity to approximate each candidate joint oscillation mode among multiple candidate joint oscillation modes with the multiple screened historical abnormal low-frequency modes, based on factors such as frequency similarity, damping difference, phase shift, and energy correlation, multiple matching results are obtained. If any of the multiple matching results has a matching degree greater than or equal to a preset matching degree threshold, the candidate joint oscillation mode passes the verification. Otherwise, the verification fails.

[0034] Furthermore, the multiple aggregated historical low-frequency anomalous mode sets are subjected to centralized representative screening to determine multiple screened historical low-frequency anomalous modes. In this embodiment, step S300 further includes: The average value of the multiple aggregated historical abnormal low-frequency modes is determined by averaging the multiple aggregated historical abnormal low-frequency modes. Using the average of the multiple aggregated historical low-frequency abnormal modes as multiple initial centers, the multiple aggregated historical low-frequency abnormal mode sets are retrieved according to a preset update bandwidth to determine multiple primary update centers; A centralized and representative comparison is performed on the plurality of primary update centers and the plurality of initial centers to determine a plurality of first-stage centers; Again, in the multiple aggregated historical abnormal low-frequency mode sets, the multiple first-stage centers are searched according to the preset update bandwidth to determine multiple secondary update centers. These centers are then compared with the multiple first-stage centers to ensure their representativeness. Based on the comparison results, the filtering continues until the preset filtering stop condition is met, thereby obtaining multiple filtered historical abnormal low-frequency modes.

[0035] Furthermore, a centralized representative comparison is performed on the plurality of primary update centers and the plurality of initial centers to determine a plurality of first-stage centers. In this embodiment, step S300 further includes: Calculate the nearest neighbor tendency coefficients of the plurality of primary update centers and the plurality of initial centers respectively to obtain the nearest neighbor tendency coefficients of the plurality of primary update centers and the nearest neighbor tendency coefficients of the plurality of initial centers; The nearest neighbor tendency coefficients of the multiple updated centers and the multiple initial centers are mapped and compared respectively. The center corresponding to the larger value in the mapping comparison result is taken as multiple first-stage centers.

[0036] Furthermore, the preset filtering stop condition is that the number of updates meets the preset maximum number of updates and / or the difference between the two nearest neighbor tropism coefficients of the mapping comparison is less than or equal to the preset difference.

[0037] In one embodiment, averaging is first performed on multiple aggregated historical anomalous low-frequency modes, i.e., the average value of each mode cluster in key characteristic dimensions, such as frequency, damping ratio, phase, and energy, is calculated to obtain multiple initial centers. Each initial center can be regarded as a representative of various typical oscillation modes under initial conditions, such as hydraulic-mechanical coupling, thermal inertial feedback, and electrical synchronous oscillation.

[0038] Subsequently, using the initial centers as a benchmark, and according to the preset update bandwidth pre-set by those skilled in the art, historical anomalous low-frequency modes that meet the bandwidth condition are retrieved within each aggregate set. One historical anomalous low-frequency mode is randomly selected from the retrieved results as multiple primary update centers. A centralized representativeness comparison is performed between the multiple primary update centers and the multiple initial centers. The nearest neighbor tendency coefficient is calculated, where each nearest neighbor tendency coefficient is the ratio of the number of historical anomalous low-frequency modes whose distance to each center is within the preset update bandwidth to the total number of modes in the corresponding aggregated historical anomalous low-frequency mode set. The larger the nearest neighbor tendency coefficient, the higher the representativeness of the corresponding center.

[0039] If the directional coefficient of the updated center is greater than that of the initial center, it means that the new center is more representative of the modal distribution, and it is retained as the first-stage center; otherwise, multiple initial centers are used as multiple first-stage centers.

[0040] Next, based on the centers from the first stage, a second round of updates and comparisons is performed on the modality set to obtain the updated centers. If the number of updates meets the preset maximum number of updates and / or the difference between the tropism coefficients of the two nearest neighbors in the mapping comparison is less than or equal to a preset difference (e.g., the change in tropism coefficient ≤ 0.01), the screening stop condition is triggered, and the clustering result is considered to have converged. The centers output at the end are the multiple screened historical abnormal low-frequency modalities.

[0041] By iteratively updating the center and comparing representativeness, high-precision aggregation and feature extraction of historical low-frequency oscillation samples are achieved, avoiding overfitting or center drift problems in traditional clustering. The resulting selected modes represent typical physical oscillation mechanisms, such as thermodynamic hysteresis-mechanical inertial resonance and electrical power angle feedback-servo oscillation, providing stable and interpretable benchmark samples for subsequent scenario consistency comparison of candidate modes.

[0042] Step S400: Use a large model to perform fault scenario-based diagnosis on the multiple abnormal low-frequency oscillation modes respectively, and obtain the target diagnosis results.

[0043] Furthermore, by utilizing a large model to perform fault scenario-based diagnosis on the multiple abnormal low-frequency oscillation modes respectively, and obtaining the target diagnosis result, step S400 of this embodiment of the application also includes: Multiple abnormal low-frequency oscillation modes and multiple sample fault diagnosis results are obtained as training data. The framework based on the feedforward neural network is trained under supervision to obtain a large model. The large model is used to identify the multiple abnormal low-frequency oscillation modes, and multiple fault diagnosis results are obtained. The target diagnostic result is obtained by summarizing the multiple fault diagnosis results.

[0044] In one embodiment, a framework based on a feedforward neural network is trained using multiple sample abnormal low-frequency oscillation modes and multiple sample fault diagnosis results as training data. A supervised training strategy, such as cross-entropy loss function and Adam optimization algorithm, is employed to learn the nonlinear mapping between complex modal features and fault types. After training, the model possesses cross-domain modal recognition and mechanism reasoning capabilities. When a new abnormal low-frequency oscillation mode is detected, the model encodes and compares its features, identifies the most similar historical modal group, and infers possible fault scenarios. For example, if the model identifies a highly coherent oscillation between hydraulic pressure and torsional vibration with a decreasing damping ratio, it outputs a hydraulic servo instability oscillation; if there is a fixed-time delay in the reverse energy transfer between boiler heat storage and power angle, it outputs electrical oscillation caused by thermal inertia feedback; if the frequency drift is synchronized with the power angle fluctuation of the power grid, it is determined to be an electromechanical composite oscillation. The multiple fault diagnosis results are then summarized to obtain the target diagnosis result.

[0045] Example 2, based on the same inventive concept as the low-frequency oscillation fault diagnosis method for thermal power units based on multi-source data analysis in the foregoing examples, as shown in the appendix. Figure 2 As shown, this application provides a low-frequency oscillation fault diagnosis system for thermal power units based on multi-source data analysis. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0046] The operation sensing module 11 is used to construct a multi-source signal acquisition network, perform multi-source unit operation sensing on the target thermal power unit through a preset monitoring window, and obtain the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence and power grid power angle sequence. The cascaded time-frequency analysis module 12 is used to traverse the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence and power grid power angle sequence to perform cascaded time-frequency analysis, and obtain multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters; The modal identification module 13 is used to perform joint frequency matching and modal identification on the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters to obtain multiple abnormal low-frequency oscillation modes. The diagnostic module 14 is used to perform fault scenario-based diagnosis on the multiple abnormal low-frequency oscillation modes using a large model to obtain the target diagnostic results.

[0047] Furthermore, the operation sensing module 11 is used to perform the following steps: High-frequency pressure sensors are deployed at the inlet and outlet of the target thermal power unit; Four sets of laser Doppler vibration meters were arranged circumferentially at the high- and intermediate-pressure rotor of the target thermal power unit. Install steam drum pressure sensors and steam flow meters on the steam drum of the target thermal power unit; The PMU device is connected to synchronously capture the grid connection point frequency and power angle of the target thermal power unit; The multi-source signal acquisition network is constructed based on a high-frequency pressure sensor, four sets of laser Doppler vibration meters, a steam drum pressure sensor, a steam flow meter, and a PMU device.

[0048] Furthermore, the modality recognition module 13 is used to perform the following steps: The governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence are traversed and multi-scale wavelet decomposition is performed to filter out high-frequency random disturbances. Modal decomposition is then performed to obtain multiple governor oil pressure pulsation eigenmode components, multiple shaft torsional vibration angular displacement eigenmode components, multiple boiler heat storage coefficient eigenmode components, and multiple power grid power angle eigenmode components. Prony analysis was performed on multiple governor hydraulic pressure pulsation intrinsic mode components, multiple shaft torsional vibration angular displacement intrinsic mode components, multiple boiler heat storage coefficient intrinsic mode components, and multiple power grid power angle intrinsic mode components to obtain multiple governor hydraulic pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters.

[0049] Furthermore, the modality recognition module 13 is used to perform the following steps: Multiple frequency oscillation interest intervals are obtained, and the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters are mapped to obtain multiple candidate joint oscillation modes. A scenario-based consistency check is performed on the multiple candidate joint oscillation modes. If the check passes, the multiple abnormal low-frequency oscillation modes are added.

[0050] Furthermore, the modality recognition module 13 is used to perform the following steps: Obtain a set of historical abnormal low-frequency modes, and perform similar aggregation on the set of historical abnormal low-frequency modes to obtain multiple aggregated sets of historical abnormal low-frequency modes; The multiple aggregated historical abnormal low-frequency modes are centrally and representatively screened to determine multiple screened historical abnormal low-frequency modes. The candidate joint oscillation modes are matched with the screening historical abnormal low-frequency modes respectively. When there is a matching result with a matching degree greater than or equal to the preset matching degree threshold, the verification is passed. If there is no matching result with a matching degree greater than or equal to the preset matching degree threshold, the verification fails.

[0051] Furthermore, the modality recognition module 13 is used to perform the following steps: The average value of the multiple aggregated historical abnormal low-frequency modes is determined by averaging the multiple aggregated historical abnormal low-frequency modes. Using the average of the multiple aggregated historical low-frequency abnormal modes as multiple initial centers, the multiple aggregated historical low-frequency abnormal mode sets are retrieved according to a preset update bandwidth to determine multiple primary update centers; A centralized and representative comparison is performed on the plurality of primary update centers and the plurality of initial centers to determine a plurality of first-stage centers; Again, in the multiple aggregated historical abnormal low-frequency mode sets, the multiple first-stage centers are searched according to the preset update bandwidth to determine multiple secondary update centers. These centers are then compared with the multiple first-stage centers to ensure their representativeness. Based on the comparison results, the filtering continues until the preset filtering stop condition is met, thereby obtaining multiple filtered historical abnormal low-frequency modes.

[0052] Furthermore, the modality recognition module 13 is used to perform the following steps: Calculate the nearest neighbor tendency coefficients of the plurality of primary update centers and the plurality of initial centers respectively to obtain the nearest neighbor tendency coefficients of the plurality of primary update centers and the nearest neighbor tendency coefficients of the plurality of initial centers; The nearest neighbor tendency coefficients of the multiple updated centers and the multiple initial centers are mapped and compared respectively. The center corresponding to the larger value in the mapping comparison result is taken as multiple first-stage centers.

[0053] Furthermore, the preset filtering stop condition is that the number of updates meets the preset maximum number of updates and / or the difference between the two nearest neighbor tropism coefficients of the mapping comparison is less than or equal to the preset difference.

[0054] Furthermore, the diagnostic module 14 is used to perform the following steps: Multiple abnormal low-frequency oscillation modes and multiple sample fault diagnosis results are obtained as training data. The framework based on the feedforward neural network is trained under supervision to obtain a large model. The large model is used to identify the multiple abnormal low-frequency oscillation modes, and multiple fault diagnosis results are obtained. The target diagnostic result is obtained by summarizing the multiple fault diagnosis results.

[0055] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0056] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0057] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis, characterized in that, The method includes: A multi-source signal acquisition network is constructed to perform multi-source unit operation sensing on the target thermal power unit through a preset monitoring window, and to obtain governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence and power grid power angle sequence. A cascaded time-frequency analysis was performed on the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence to obtain multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters. By performing joint frequency matching and mode identification on the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters, multiple abnormal low-frequency oscillation modes are obtained. The large model is used to perform fault scenario-based diagnosis on the multiple abnormal low-frequency oscillation modes to obtain the target diagnosis results.

2. The method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis as described in claim 1, characterized in that, Constructing a multi-source signal acquisition network includes: High-frequency pressure sensors are deployed at the inlet and outlet of the target thermal power unit; Four sets of laser Doppler vibration meters were arranged circumferentially at the high- and intermediate-pressure rotor of the target thermal power unit. Install steam drum pressure sensors and steam flow meters on the steam drum of the target thermal power unit; The PMU device is connected to synchronously capture the grid connection point frequency and power angle of the target thermal power unit; The multi-source signal acquisition network is constructed based on a high-frequency pressure sensor, four sets of laser Doppler vibration meters, a steam drum pressure sensor, a steam flow meter, and a PMU device.

3. The method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis as described in claim 1, characterized in that, A cascaded time-frequency analysis was performed on the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence to obtain multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters, including: The governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence are traversed and multi-scale wavelet decomposition is performed to filter out high-frequency random disturbances. Modal decomposition is then performed to obtain multiple governor oil pressure pulsation eigenmode components, multiple shaft torsional vibration angular displacement eigenmode components, multiple boiler heat storage coefficient eigenmode components, and multiple power grid power angle eigenmode components. Prony analysis was performed on multiple governor hydraulic pressure pulsation intrinsic mode components, multiple shaft torsional vibration angular displacement intrinsic mode components, multiple boiler heat storage coefficient intrinsic mode components, and multiple power grid power angle intrinsic mode components to obtain multiple governor hydraulic pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters.

4. The method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis as described in claim 1, characterized in that, Joint frequency matching and mode identification were performed on the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters, and multiple power grid power angle oscillation parameters to obtain multiple abnormal low-frequency oscillation modes, including: Multiple frequency oscillation interest intervals are obtained, and the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters are mapped to obtain multiple candidate joint oscillation modes. A scenario-based consistency check is performed on the multiple candidate joint oscillation modes. If the check passes, the multiple abnormal low-frequency oscillation modes are added.

5. The method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis as described in claim 4, characterized in that, A scenario-based consistency check is performed on the multiple candidate joint oscillation modes. If the check passes, the modes are added to the multiple abnormal low-frequency modes, including: Obtain a set of historical abnormal low-frequency modes, and perform similar aggregation on the set of historical abnormal low-frequency modes to obtain multiple aggregated sets of historical abnormal low-frequency modes; The multiple aggregated historical abnormal low-frequency modes are centrally and representatively screened to determine multiple screened historical abnormal low-frequency modes. The candidate joint oscillation modes are matched with the screening historical abnormal low-frequency modes respectively. When there is a matching result with a matching degree greater than or equal to the preset matching degree threshold, the verification is passed. If there is no matching result with a matching degree greater than or equal to the preset matching degree threshold, the verification fails.

6. The method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis as described in claim 5, characterized in that, The multiple aggregated historical low-frequency anomalous mode sets are centrally and representatively screened to determine multiple screened historical low-frequency anomalous modes, including: The average value of the multiple aggregated historical abnormal low-frequency modes is determined by averaging the multiple aggregated historical abnormal low-frequency modes. Using the average of the multiple aggregated historical low-frequency abnormal modes as multiple initial centers, the multiple aggregated historical low-frequency abnormal mode sets are retrieved according to a preset update bandwidth to determine multiple primary update centers; A centralized and representative comparison is performed on the plurality of primary update centers and the plurality of initial centers to determine a plurality of first-stage centers; Again, in the multiple aggregated historical abnormal low-frequency mode sets, the multiple first-stage centers are searched according to the preset update bandwidth to determine multiple secondary update centers. These centers are then compared with the multiple first-stage centers to ensure their representativeness. Based on the comparison results, the filtering continues until the preset filtering stop condition is met, thereby obtaining multiple filtered historical abnormal low-frequency modes.

7. The method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis as described in claim 6, characterized in that, A centralized representative comparison is performed on the plurality of primary update centers and the plurality of initial centers to determine a plurality of first-stage centers, including: Calculate the nearest neighbor tendency coefficients of the plurality of primary update centers and the plurality of initial centers respectively to obtain the nearest neighbor tendency coefficients of the plurality of primary update centers and the nearest neighbor tendency coefficients of the plurality of initial centers; The nearest neighbor tendency coefficients of the multiple updated centers and the multiple initial centers are mapped and compared respectively. The center corresponding to the larger value in the mapping comparison result is taken as multiple first-stage centers.

8. The method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis as described in claim 7, characterized in that, The preset filtering stop condition is that the number of updates meets the preset maximum number of updates and / or the difference between the two nearest neighbor tropism coefficients of the mapping comparison is less than or equal to the preset difference.

9. The method for diagnosing low-frequency oscillation faults in thermal power units based on multi-source data analysis as described in claim 1, characterized in that, Using a large model, fault scenario-based diagnosis is performed on the multiple abnormal low-frequency oscillation modes to obtain target diagnostic results, including: Multiple abnormal low-frequency oscillation modes and multiple sample fault diagnosis results are obtained as training data. The framework based on the feedforward neural network is trained under supervision to obtain a large model. The large model is used to identify the multiple abnormal low-frequency oscillation modes, and multiple fault diagnosis results are obtained. The target diagnostic result is obtained by summarizing the multiple fault diagnosis results.

10. A low-frequency oscillation fault diagnosis system for thermal power units based on multi-source data analysis, characterized in that, The system is used to implement the low-frequency oscillation fault diagnosis method for thermal power units based on multi-source data analysis as described in any one of claims 1-9, and the system comprises: The operation sensing module is used to construct a multi-source signal acquisition network, perform multi-source unit operation sensing on the target thermal power unit within a preset monitoring window, and obtain the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence, and power grid power angle sequence. The cascaded time-frequency analysis module is used to traverse the governor oil pressure pulsation sequence, shaft torsional vibration angular displacement sequence, boiler heat storage coefficient sequence and power grid power angle sequence to perform cascaded time-frequency analysis, and obtain multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters; The modal identification module is used to perform joint frequency matching and modal identification on the multiple governor oil pressure pulsation oscillation parameters, multiple shaft torsional vibration angular displacement oscillation parameters, multiple boiler heat storage coefficient oscillation parameters and multiple power grid power angle oscillation parameters to obtain multiple abnormal low-frequency oscillation modes. The diagnostic module is used to perform fault scenario-based diagnosis on the multiple abnormal low-frequency oscillation modes using a large model to obtain the target diagnostic results.

Images